Mini Review - (2016) Volume 2, Issue 2
James B Hittner*
Department of Psychology, College of Charleston, USA
*Corresponding Author:
James B Hittner
Department of Psychology,
College of Charleston
66 George Street Charleston SC 29424, USA.
Tel: 8439536734
Fax: 8439537151
E-mail: hittnerj@cofc.edu
Received Date: March 15, 2016; Accepted Date: March 31, 2016; Published Date: April 07, 2016
Citation: Hittner JB. On Correctly Adjusting the Squared Multiple Correlation Coefficient in Linear Regression: Effect Size Estimation and Significance Testing with Application to Substance Abuse Research. J Drug Abuse. 2016, 2:2.
Linear regression analysis is ubiquitous in many areas of scholarly inquiry, including substance abuse research. In linear regression it is common practice to test whether the squared multiple correlation coefficient, R2, differs significantly from zero. However, this test is misleading because the expected value of R2 is not zero under the null hypothesis. In this brief methodological note I discuss the implications of this realization for calculating and interpreting the squared multiple correlation coefficient, R2. In addition, I discuss and offer freely available software that calculates the expected value of R2 under the null hypothesis that ρ-the population value of the multiple correlation coefficient-equals zero, an adjusted R2 value and effect size measure that both take into account the expected value of R2, and an F statistic that tests the significance of difference between the obtained R2 and the expected value of R2 under the null hypothesis that ρ=0.
Keywords
Multiple correlation; Regression; Effect size; Hypothesis testing; Computer program; SAS
Suppose that a researcher is interested in predicting cocaine use frequency in adults from a number of putative risk factors such as sensation seeking tendencies, level of psychosocial stress, frequency of other illicit drug use, and number of negative life events. A commonly used statistical approach to modeling such data is linear multiple regression analysis, and when using regression it is standard practice to examine whether the squared multiple correlation coefficient, R2 (i.e., the proportion of variance in the outcome variable accounted for by the predictors) is statistically significant. The intent of such a test is to determine whether R2 differs significantly from zero, and the null hypothesis may be stated as Ho: ρ2=0, where ρ2 represents the population value (parameter) for the squared multiple correlation coefficient. Although this test is widely used, it is misleading because the expected value of R2 is not zero when ρ=0 (where ρ represents the population value for the multiple correlation coefficient). Rather, as Morrison [1] pointed out, the expected value, or expected long-run mean, of R2 is equal to p/n-1, where p is the number of predictor variables and n is the sample size. The implication of this equation is that R2 should be examined in relation to the expected value of R2, E(R2), because the latter quantity is the value of R2 that can be expected simply “by chance.” In light of this realization, it seems more appropriate for researchers to test the null hypothesis, Ho: ρ2=ρo 2, where ρo 2=E(R2). Unfortunately, commonly used statistical software programs such as IBM SPSS, Minitab, and SAS do not implement this adjusted hypothesis test whereby R2 is tested against the expected value of R2.
Huberty [2] recognized the importance of testing R2 against its expected value and he proposed an adjusted R2 index that takes into account the value of E(R2). This adjusted R2 index “corrects” the obtained squared multiple correlation coefficient by explicitly incorporating the expected value of R2. Huberty [2] also presented an effect size measure for linear multiple regression studies that is calculated by subtracting E(R2) from Huberty’s adjusted R2 index. Darlington [3] gave an F statistic for testing the null hypothesis that R2 equals the expected value of R2 (i.e., Ho: ρ2=ρo 2). To calculate the aforementioned statistical quantities, I wrote a user-friendly computer program called DARHUB (short for Darlington and Huberty statistics). DARHUB calculates the expected value of R2, Darlington’s F statistic for testing the null hypothesis that ρ2=ρo 2, the observed probability value for Darlington’s F, Huberty’s adjusted R2 index, and Huberty’s effect size measure. The user specifies only the squared multiple correlation coefficient, R2, from a linear regression analysis (i.e., the R2 obtained by performing a regression analysis using SPSS, Minitab, SAS, etc.), the number of predictor variables, p, and the sample size, n. DARHUB is written in SAS data step language [4], and SAS is compatible with PC, Mac and Linux workstations. User documentation is included in the program listing. Table 1 contains sample input, the full program listing (the SAS data step code), and output from the analysis (Table 1). The DARHUB program and accompanying ReadMe file can be downloaded from the following academic website: https://psychology.cofc. edu/personal-pages/james-hittner,-phd.---software-page.php.
Input and program listing: d ata R squared; *** Below, specify the sample size "n", the number of predictor variables "p", and the squared multiple correlation coefficient "Rsq" ***; n=60; p=5; Rsq=0.31; ***********************************************************************; ERsq=((p / (n-1))); a=((ERsq / (1-ERsq))); ******************* Darlington's F below ***********************************; Dar_F=((Rsq / (1-Rsq)) * ((n-p-1) / p)) / (1 + (a * ((n-1) / p))); ***********************************************************************; v=(((ERsq * (n-1)) + p))**2 / (((ERsq * (n-1) * (ERsq + 2)) + p)); Dar_p=((1 - probf(Dar_F, v, n-p-1))); ************* Huberty's adjusted Rsq value below ****************************; HubadRsq=((Rsq - ERsq) / (1 - ERsq)); ************* Huberty's effect size index below *****************************; Hubefsiz=(HubadRsq - ERsq); ***********************************************************************; proc print noobs; var Dar_F Dar_p ERsq HubadRsq Hubefsiz; options nodate; options nocenter; title1 'NOTE: Dar_F=F test reported in Darlington (1990) for testing the'; title2 'null hypothesis that R-squared equals the expected value of R-squared.'; title3 'Dar_p=Probability value for Dar_F.'; title4 'ERsq=The expected value, or long-run mean, of R-squared under the null'; title5 'hypothesis that rho=0.'; title6 'HubadRsq=Adjusted R-squared value presented in Huberty (1994). Expressed'; title7 'as a proportional-reduction-in-error or improvement-over-chance statistic.'; title8 'Hubefsiz=Effect size index presented in Huberty (1994). Calculated by'; title9 "subtracting the expected value of R-squared from Huberty's adjusted"; title10 'R-squared value.'; run; |
Output:
NOTE: Dar_F=F test reported in Darlington (1990) for testing the |
null hypothesis that R-squared equals the expected value of R-squared. |
Dar_p=Probability value for Dar_F. |
ERsq=The expected value, or long-run mean, of R-squared under the null |
hypothesis that rho=0. |
HubadRsq=Adjusted R-squared value presented in Huberty [2]. Expressed |
as a proportional-reduction-in-error or improvement-over-chance statistic. |
Hubefsiz=Effect size index presented in Huberty [2]. Calculated by |
subtracting the expected value of R-squared from Huberty's adjusted |
R-squared value. |
Dar_F | Dar_p | ERsq | HubadRsq | Hubefsiz |
---|---|---|---|---|
2.31874 | 0.042009 | 0.084746 | 0.24611 | 0.16137 |
Table 1: The DARHUB SAS program with sample input and output.